Optimize the aperture photometry operation

[jnation3406]

The aperture photometry was killing the production pods with OOM when trying to do it with 10 QHY full frame images. This PR is mostly Claude implementing a few optimizations to prevent this:

1. Decouple measuring and ranking candidate stars so first we measure stars for each image one by one, closing the image when we are done. Then rank all the candidates
2. Don't upcast to data to float64 in aperture_light_curve.py - keep it as float32
3. Downsample the diagnostic jpegs created to a max dimension of 1500x1500 - a reasonable size to be displayed on the screen.

I also threw in some optimization of the wcs and source catalog analysis task, to not load the unused pixel data into those tasks.

Claudes description of changes:
aperture_light_curve.py — the core restructure. generate_light_curve now takes local FITS paths instead of eagerly-loaded handlers, and runs in three phases: a metadata pass that validates frames reading only the SCI header and CAT table (no pixels), the candidate catalog build (headers only), then a streamed pixel pass (_measure_frame_pixels) that loads one frame as float32, measures the target plus every candidate on it, captures a downsampled preview, and releases the pixels before the next frame loads. FrameContext no longer has an .image field, and CAT rows are trimmed to the six columns the pipeline actually reads.

comparison_stars.py — the candidate-outer/frame-inner measurement loop (which forced all frames to coexist) is gone. candidate_stars_from_catalog builds candidates up front, the pixel pass measures them frame-by-frame, and select_comparison_stars is now pure math over the collected measurements, with identical drop/rank semantics.

photometry_diagnostics.py — overlays are drawn on a block-mean-downsampled FramePreview (capped at your 2000 px) captured during the pixel pass. That removes the GB-scale full-frame render temporaries and shrinks each frame's base64 JPEG in the operation output (and Redis, and the API response) from tens of MB to well under 1 MB. Small frames like the 80×80 test fixtures are untouched (scale 1.0).

aperture_photometry.py — operate() resolves inputs to file-cache paths via FileCache directly; no more list of eager InputDataHandlers each holding a decompressed frame.

centroiding.py:260 — removed the dtype=float upcast that would otherwise have silently copied the whole frame to float64 on every centroid call (once per candidate per frame).

[sfoale]

Nicely done, Claude.

[capetillo]

I think it looks okay but I'm confused if the ranking changed. Anyway, good job Jon & Fable